observations with missing values in newdata might not have gone with the majority in predict(, newdata) when no surrogate split was applicable. Spotted and fixed by Vik .
